Mismatch between code and dataset contents

Hello,

I tried to test the provided code with the provided dataset, code and checkpoint as described in the README:

Used dataset is : $ wget https://www.dropbox.com/s/y5kptmgxcngesp7/RADIal.zip
Used code is : https://github.com/valeoai/RADIal/tree/main/FFTRadNet
Used checkpoint is : https://www.dropbox.com/s/fnk92jmvd57ijwp/FFTRadNet_RA_192_56_epoch78_loss_172.8239_AP_0.9813.zip

There is a compatibility issue between line 71 of file FFTRadNet/dataset/dataset.py) and the expected file extension (npy expected in the code but npz provided)

I found a quick fix by having a look at RADIal/loader/dataset.py) file, line 60. It consists in referring to npz file, arr_0 file:

radar_name = os.path.join(self.root_dir,'radar_FFT',"fft_{:06d}.npz".format(sample_id))
input = np.load(radar_name,allow_pickle=True)['arr_0']

With this fix, I am now able to launch 1-Train.py, 2-Test.py and 3-Evaluation.py scripts but more serious problems happened:

$ python 2-Test.py --config config/config_FFTRadNet_192_56.json --checkpoint FFTRadNet_RA_192_56_epoch78_loss_172.8239_AP_0.9813.pth Launches an opencv window, camera images are updated, detected cars are always missed (bounding box always at the center of the image) whereas fft spectrum and freespace segmentation parts do not change over time. The the application crashed after several samples being displayed:
```
ASSERT: "false" in file qasciikey.cpp, line 501
Aborted (core dumped)
```
$ python 3-Evaluation.py --config config/config_FFTRadNet_192_56.json --checkpoint FFTRadNet_RA_192_56_epoch78_loss_172.8239_AP_0.9813.pth result in very small scores for a pretrained model:

===========  Dataset  ==================:
      Mode: sequence
      Training: 6118
      Validation: 786
      Test: 744

===========  Loading the model ==================:
===========  Running the evaluation ==================:
Generating Predictions...
743/744 [==================>.] - ETA: 0s------- Detection Scores ------------
  mAP: 0.002538829151732378
  mAR: 0.002457002457002457
  F1 score: 0.002497245684906354
------- Regression Errors------------
  Range Error: 0.08149683475494385 m
  Angle Error: 0.2016620635986328 degree
------- Freespace Scores ------------
  mIoU 63.01938749672122 %

Test0 Test1

Do you have an explanation and a suggestion for using FFTRadNet properly ?

Best regards, Nicolas

valeoai / RADIal

Mismatch between code and dataset contents #10